Newsgroups: comp.arch
Path: utzoo!utgpu!watserv1!watdragon!watsol.waterloo.edu!tbray
From: tbray@watsol.waterloo.edu (Tim Bray)
Subject: Re: Networking for Distributed Computing
Message-ID: <1991Apr6.012655.9480@watdragon.waterloo.edu>
Sender: news@watdragon.waterloo.edu (News Owner)
Organization: University of Waterloo
References: <1991Apr5.182853.20728@hubcap.clemson.edu>
Date: Sat, 6 Apr 1991 01:26:55 GMT
Lines: 26

mccalpin@perelandra.cms.udel.edu (John D. McCalpin) writes:
 Unfortunately, the most commonly available networking
 option (ethernet) uses a broadcast approach, which is definitely
 sub-optimal for the communications needs of many "natural"
 parallel distributed algorithms.
 ...proposed complex SCSI message-passer...
 So what is wrong with this idea?
 done with a SCSI interface or how hard the implementation of a
 buffered FIFO would be?

What's wrong with this idea: first you have to prove that when you build
this system, you have a bottleneck at the Ethernet.  Then you have to
prove you can't fix it by just dropping a same-but-faster FDDI spine or
suchlike.

As for me, when I first saw the Ethernet protocol, I said: mickey mouse -
that'll never fly for real work.  When I heard about people wanting to use
TCP/IP for local-area-networking, I said: that protocol will blow Ethernet
out of the water in about 10 minutes.  Sigh, 0 for 2, and another lesson in
the futility of intuition in predicting performance bottlenecks.  My own data
point: I've worked on a lot of different distributed environments, and 95% of
the time, you run out of CPU, or filesystem bandwidth, or context switches,
or something, before your Ethernet runs out of gas.

Of course, you could be right.
Tim Bray, Open Text Systems
